Research On Algorithms Of Fish Recognition And Detection Based On Deep Learning

The ocean is the largest ecosystem on the planet and contains very rich species resources.my country’s ocean area is 3 million square kilometers,but fishery resources account for a very large proportion of ocean resources,and it has shown a gradual decline in recent years.Fish classification and identification have broad application prospects in the fields of fishery resource research,scientific promotion of fish knowledge,aquaculture processing,and rare species protection.Real-time monitoring of the seabed environment in the surrounding sea area,analysis and study of the distribution of representative fish in the sea area and living habits are of great significance for the sustainable development of fishery resources and the detection of unknown biological resources.This article mainly conducts research through the following aspects:(1)A total of 908 pictures of ten species of bonito,turbot,needlefish,black snapper,yellowfin snapper,goldfish,perch,greenfin,silver pomfret,and mullet were collected,and the images were classified and identified.A fish image recognition algorithm based on the Keras deep learning framework is proposed.The ResNet50 is the basic network framework,and the confusion matrix is used to judge the pros and cons of the network model.It is intuitive through the precision,recall,and f1-score indicators in the confusion matrix.Show the accuracy of the model.Using this model to classify and recognize 10 species of fish,the results show that the correct rate of recognition reached 93.33%.(2)For turbot,yellowfin sea bream,goldfish,and mullet,the number of data sets was increased.A total of 1123 images were collected.For each fish,42 images were set in the verification set,and the same preprocessing operation was adopted.Using the PyTorch framework structure and the ResNet50 network model,different algorithms are used to classify and recognize them,and the model is continuously optimized.The four fishes are trained and learned,and the accuracy of the test is more than 96%.(3)The GUI visual interface was developed with PyQt5,and the test operation was performed through the selection and prediction function buttons in the interface.The test results showed that the actual category was consistent with the predicted category.The realtime tracking and detection of underwater targets was done using the DSOD framework.The test results showed that small objects were added.The detection accuracy of this model is maintained,while the detection rate of the original model is maintained,and the detection result reaches the expectation.